Landslide Susceptibility Prediction and Driving Force Analysis Integrating Machine Learning and Spatial Factor Optimization: A Case Study in the Guanyinyan Hydropower Station Reservoir Area

ABSTRACT Landslides are complex geological hazards driven by the interaction of multiple factors, exhibiting significant spatial heterogeneity. Although machine learning has made notable progress in landslide susceptibility prediction, most models still rely on expert knowledge or fixed rules for feature discretization, limiting their adaptability across scales and spatial expressiveness. To address the aforementioned issues, this study introduces the optimal parameters‐based geographical detector (OPGD) to determine the optimal classification intervals for environmental factors, which are then coupled with six machine learning models, including logistic regression (LR), random forest (RF), support vector machine (SVM), Naive Bayes (NB), K‐nearest neighbors (KNN), and multilayer perceptron (MLP) to predict landslide susceptibility in the Guanyinyan hydropower station reservoir area. Additionally, the Shapley Additive Explanations (SHAP) is employed to further identify the key driving factors and their nonlinear response characteristics. The results show that (1) the optimal classification of 15 factors into 7–9 categories yields the highest spatial heterogeneity explanatory power, significantly improving the representation of landslide spatial patterns; (2) RF and SVM models outperform others, with training AUC values above 0.90 and high‐risk zones covering 24.87% and 23.21% of the study area, respectively; (3) normalized difference vegetation index (NDVI), human footprint index (HFI), distance to waters (DTWs), and elevation (ELE) emerge as the dominant drivers. NDVI is negatively associated with landslide risk, while HFI, DTW, and ELE show positive associations, revealing a compound mechanism shaped by topographic, ecological, and anthropogenic interactions. The framework developed in this study balances the objectivity of factor representation, the stability of model prediction, and the interpretability of the underlying mechanisms, effectively supporting spatial identification of landslide risks in similar regions.